186 | Background check for AI Agents, AI work force benefits challenged, AI risks mounting, and more important AI news for the week ending on May 2nd 2025

25:16And once you start adding the military aspect of this, it's a very, very slippery slope of we want to have something that the other guys won't have or have it first. And the distance between that and Cybernet and a catastrophic global war style Terminator is not that far. And now two rapid fire items. We're going to start with Microsoft. A lot of things are happening in Microsoft and there's a lot of interesting news from them this week. So the first piece of news is actually not that much news. It's just more updates on the same thing. And that's the growing tension between Microsoft and OpenAI and more specifically between Sam Altman and Satya Nadella.

25:47So as you all probably know, the reason OpenAI became what they are is because of$14 billion investment by Microsoft that took them from being just a small research group to the$300 billion valuation behemoth they are right now. And yet, this relationship has been growing further and further apart between OpenAI claiming that Microsoft is not providing them enough access and Microsoft establishing their own internal research group and bringing in Suleiman to run that group. So that aspect of it. On the flip side, Microsoft are claiming that OpenAI are not providing them access to the latest and greatest, and they're limiting their access to different types of AI that they're releasing on their own.

26:23So not the best relationship. There's this particular piece of news is sharing also that they went down from texting to each other five to six times a day to talking about once a week, which is definitely showing that beyond the business relationship, the personal relationship is not at the same level as it was before. Now, while that's great gossip news, I think what it shows is it shows that both companies are growing beyond this initial relationships. And it makes sense for both companies to have other options, right? It allows OpenAI to get access to additional compute from different providers for additional directions of research that may or may not benefit Microsoft.

26:57And it allows Microsoft to diversify its offering to its users. By the way, since we mentioned Satya, he mentioned in an interview this week that 20 to 30 % of Microsoft code is now AI written. Now, there's no way for me to imagine how much code Microsoft generates in a week, but that's probably a huge amount because that's what they do. And so if 20 to 30 % of that is currently generated by AI, it is absolutely astonishing to me that at the enterprise level, one of the largest companies in the world and definitely one of the largest generators of code in the world, 20 to 30 % of code is already generated by AI, connecting us back to our point of training and setup in the organization, both from a strategic perspective and as well as from a tactical perspective of how impactful AI already is.

27:43Now, to connect to our previous point of Microsoft relationship and them offering additional options for their clients, there is an interesting rumor that is saying that Microsoft is getting ready to provide XAI's Grok model on Azure AI Foundry, which is their AI platform on Azure. Now, the reason this is interesting is not because it's the first model beyond OpenAI, because they're allowing other models to run there already, but because Elon Musk is currently the number one hater of Sam Altman, which is obviously the CEO of OpenAI. And so from a personal perspective, as well as from the fact that Elon Musk started XAI to compete with OpenAI and just show them that he can do it better, just creates a very complex scenario from an emotional perspective where Microsoft is going to offer OpenAI's largest competitor on their platform, in addition to obviously offering OpenAI.

28:38Now, this has not been finalized yet, but if it is finalized, it means that XAI will be available on Azure for companies to choose instead of choosing OpenAI or in addition to choosing OpenAI and integrate it into anything that they want. Now, the reality is I don't think that Microsoft has a choice because if the other two platforms, Google Cloud and AWS, are going to provide multiple options to their clients, I think Azure will have to do the same. And so Microsoft doesn't have much of a choice but to add more and more models to its platform, even if they are competing with their number one partner in AI.

29:09That being said, Microsoft made it very, very clear that they will provide hosting and access to the models, but they will not provide them compute for training Grok. I don't think that's a problem because XAI raised a ridiculous amount of money recently, and they're in the process of potentially raising even more, and they are building their own infrastructure to train their models. And so from that perspective, at least Microsoft is not having a conflict of interest with the compute they're providing to. An interesting piece of news from Microsoft this week is that they are finally releasing the controversial recall feature.

29:38So if you remember, they have announced this about a year ago, showing that Copilot Plus PCs will have the capability to basically recall anything you do on the computer by continuously taking screenshots of what's happening on the screen and being able to search through that. That feature was not released because of significant privacy and security concern that was raised when it was announced. So this feature is being released and is going to be available starting immediately on Copilot plus PCs. In addition, they're adding a new feature called Click2Do, which allows the AI to take actions on the screen based on what's happening on the screen right now.

30:14And also advanced search capabilities like the ability to describe what is in a file in simple English and have the operating system go and find that file for you based on your description versus based on keywords. Now, different than the original idea, recall is going to be an opt-in option, meaning by default, it's going to be disabled. And to make it even safer, all the data is encrypted and processed on device, which means the data doesn't go anywhere. There is still a serious security concern. If somebody gets access to your computer, their ability to find information on it becomes a lot easier.

30:48So what's the bottom line? I don't think there is a bottom line. I think tools like that, and I'm going back to an earlier topic from this episode, which is the connectivity of Claude into multiple company systems. And the same thing will happen with all the other tools as well. I just think we're going into a world where you'll be able to ask a question about anything and get an answer that is actually based on actual information that is going to pull from everything it knows, which will be everything it has access to, including everything you did on your computer, everything in your company network, every tool that it is connected to.

31:16And I think it's going to create a complete nightmare for IT managers and CISOs in companies, chief information security officers, because it is going to dramatically change the way we work right now when it's only humans accessing data, where we already have infrastructure to actually define who has access to what. That is going to have to change and it has to change very, very quickly. Another big release from Microsoft this week, they just released a new family of open source models called in their Fi family. And their top one in this new family is Microsoft Fi4 Reasoning Plus model, which has only 14 billion parameters, which is actually relatively small, but it matches OpenAI's O3 Mini and DeepSeek's R1 on several different benchmarks.

32:00Now, the way they trained this model, that again is significantly smaller and yet achieves similar results, they trained it through a process called distillation. And the interesting thing is that they use DeepSeek R1 to train PHY4. So they deep-seeked DeepSeek. So DeepSeek did exactly the same process, most likely on OpenAI 4.0 and O1 in order to train their R1 model. And now Microsoft did the same exact thing using the R1 model to train their own model. And they're not even trying to hide that. Microsoft spokesperson himself said, using distillation, reinforcement learning, and high quality data, these models balance size and performance.

32:38Now, all the three models that just released are already available on Azure AI Foundry and on Hugging Face for anybody who wants to use them for more or less anything. Again, cheap, reliable, and excelling at specific tasks. And speaking of new models and interesting competition in the AI race, we'll switch to Meta. So we told you that Meta released Llama 4 just a couple of weeks ago. Well, Llama 4 API is now running on Cerebra's infrastructure and it achieves 2 ,648 tokens per second running Lama for Scout. That's 18 times faster than OpenAI's ChatGPT that only has 130 tokens per second through its API.

33:19And it's 20 times faster than DeepSeek with 25 tokens per second. And this is based on benchmark done by third parties. So if you want to run a super fast AI through the API, Right now, Lama's Cerebrus partnership is by far the fastest. Now, to explain a little bit what's happening here, Cerebrus, as well as Samba Nova and Grok with a Q, G-R-O-Q, are three companies that have been developing a completely different infrastructure from GPUs to be optimized for token generation. So they're not good for training new models, but they are significantly more efficient than GPUs in generating tokens. And again, to put things in perspective, while this particular parameter with Cerebrus and Lama generates over 2 ,600 tokens a second versus 130 or 25 tokens from the competition.

34:07Using SambaNova, you can generate almost 750 tokens per second and using Grok about 600 tokens per second. And I'm sure each and every one of these companies will have slightly different parameters on different models. But what it shows is it shows that there is new hardware there that can dramatically accelerate the AI responses. I can tell you that in some of the tools that I'm using, Lama with Grok, the software company, so GROQ, and the speeds in which I'm getting results are insane. You literally get full pages within splits of a second, which is nothing like we've been used to. And I think that's a step in the right direction because it's going to save electricity, it's going to save power, and it's going to give us more output for less cost, less time, and less power that it is consuming.

34:45Now, in addition to running really fast, LAMA 4 models also cost significantly less than running, let's say, OpenAI. In some cases, an order of magnitude less than running OpenAI, depending on the specific models you pick. But not everything is rainbows and butterflies in the Meta AI universe because this week, a Wall Street Journal investigation found that Meta's AI chatbots on Facebook, Instagram, and WhatsApp engaged in sexually explicit romantic role play with underaged users. Now, in Meta's response to this, they're saying that this content is actually negligible and that it made up less than 0.02 % of the engagement with their chatbot.

35:26Specific tests by Wall Street Journal shows that this sex fantasy talk with minors is definitely a possibility. That's just part of the problem. The other part of the problem is these chatbots are mimicking non-stars like John Cena and Kristen Bell and Judi Dench, as well as characters from Frozen. And that's not necessarily with approval. And in the Disney case, they stated exactly the opposite. They stated that Meta has no rights to use any of their characters for anything, and definitely not for sexting with minors. Meta themselves says the Wall Street Journal testing were manipulative and trying on purpose to achieve these goals, but what if the users themselves take the same kind of action?

36:04And that kind of shows how many loopholes there are in this whole AI universe that we're running into, not just stepping into, and how many things that we didn't even think about that we need to find ways to prevent in order to keep our kids and our universe safe. But back to positive news from Mera. Mera has been on fire with releasing new capabilities for Ray-Ban. Some of them were shared with you already in previous episodes. Some of them are new. So on a quick recap on new capabilities of Mera's Ray-Ban, they are now capable to do real-time translation of French, Italian, Spanish, and English back and forth while working offline.

36:41Meaning you can do this without internet access in countries that speak this language, which I find it to be an incredible feature for connecting people around the world. They also relaunched the Meta AI app for enhanced functionality, including allowing hands-free AI interactions, such as asking questions about what you see on multiple aspects of your doing, whether for personal use or business use. And Meta are planning to launch an advanced Ray-Ban glasses that is codenamed Hypernova that will have a display on the screen, meaning you won't just be able to see what you're seeing and talk to you.

37:15It will be able to actually provide you text information and it will be able to respond to hand gestures in order to activate and manipulate different applications and things you're connected to. The price point is expected to be between$1 ,000 and$1 ,400, so not cheap, but for a leading AI wearable device, that may not be that bad. And they're planning to release this by the end of this year. Now, the group that developed these glasses, Meta's Reality Labs, has been reported to be at a$4.2 billion operating loss in Q1 of 2025. That's despite generating$412 million in sales, which is the only company that generates any AR or VR hardware that is actually worth mentioning in that scale of revenue.

37:57To make this even more extreme, Reality Labs has racked up over $60 billion in losses since being established in 2020, but they were working mostly on metaverse infrastructure and things like that, which did not really materialize. And I think their current focus on these smart glasses are the direction to go. I'll be extremely surprised, as I mentioned in previous episodes as well, if we don't see the other big companies such as OpenAI and definitely Apple and Google coming up with similar solutions that are going to be wearable and most likely glasses because the form factor just makes sense that we'll have AI enabled capabilities, which means going back to the world we're going to live in, everybody will record everybody.

38:37Everybody will be able to analyze everything they see, including your emotion, including your actions, including your store, and including everything else in real time. And that raises so many concerns, both from a privacy perspective, as well as from data security perspective and so on, that nobody is ready for. And yet, again, it's already here. It's not just coming because you can buy these Ray-Ban glasses right now. Now, I know you're shocked. It's been more than a few minutes in this episode and we haven't talked about OpenAI yet. So here we are with our OpenAI segment. Singapore Airlines just announced their partnership with OpenAI to integrate OpenAI's platforms into multiple aspects of their business, including customer support and operations, making them the first major airline to partner with OpenAI.

39:20Their goal is obviously to deliver faster, more personalized responses in their current existing virtual assistant, Chris, but also to enable their in-house employees to optimize complex tasks like crew scheduling by using AI capabilities and integrating them so they can look at previous scenarios and get insights from that in order to make better decisions for everything in the operational side of the airline. I find this a great use case for AI. As somebody who travels a lot, I would love to see airlines being more efficient and solve problems in a more effective way by using AI looking at past use cases.

39:53The bigger piece of news from OpenAI this week was actually not positive. OpenAI released a new version of GPT-4.0 just a week ago. I shared that with you on the episode. But that version was way too flattering and agreeable in a level that was actually not comfortable and in many cases harmful to many users. Sam Altman even acknowledged that and even called it a psychophant and annoying version of ChatGPT. He also shared that they rolled back this entire version and that they're working to fix the problem. Some of it was just annoying. It was just agreeing with everything you said in too much of an extreme way and not providing valuable information when it was against what the user was asking.

40:32But in some cases, it was actually dangerous with one user citing a response endorsing a dangerous decision to stop psychotherapy and medication from a specific individual. As I mentioned, OpenAI admitted that it was a mistake and they claimed that the reason is that this new model was too much heavily focused on short-term feedback versus the broader picture, and they pledged to refine this and release a better model, but for now, it was rolled back. This comes to show you how much even the leading labs do not really know how these models work, and they do not know how one change they make is going to impact other aspects of these models.

41:11Tying it back to the risks that we discussed in the beginning shows you how risky this whole game is because the labs themselves are not 100 % sure what's going to happen when to make changes to the point that they're releasing these models to the public without being aware of an issue like this existing. Another interesting piece of news from OpenAI, we actually shared hints to that in the previous episode, is that Chachupiti now is going to provide shopping recommendations on its platform in a gallery style view. Now, OpenAI stated, and I'm quoting, product results are chosen independently and are not ads, basically meaning they are trying to find the best buying opportunity for you based on pricing reviews and your history, meaning its own memory, and it does not have any affiliate kickbacks or monetization, at least as of right now.

41:56This is running out for pro users and plus users. And as I mentioned, it integrates the memory feature, which means it will know your preferences and will help you to find stuff that's relevant to you. This is just a first step into a completely new internet that nobody is ready for and know exactly where it's going. But if you think about it, it means that you will not browse websites, meaning everything we know about the internet and about UI design and about psychology of users in order to get them to buy, shop, or take specific action is going to change because humans are not going to visit websites.

42:31Agents are going to visit websites and aggregate information for them. And it means that the entire structure of the internet and the entire monetization of the internet has to change in order for this to be a successful transition. What does that mean? It means that there's a huge opportunity right now for companies to figure this out first while the transition is happening and learn how to build websites that are optimized for agents while in the meanwhile, keeping the front end of it as well. But then when more and more of the traffic is going to be agent-based, they will win significant market share over companies who do not take that step.

43:04And the last piece of news is not really OpenAI, but it's related to Sam Altman. So Sam Altman's interesting project with its iris-reading orb devices just launched in multiple locations around the world. The launch wasn't completely successful because there were a lot of bugs and issues with the device. But the project that Sam Altman is one of its co-founders, and he co-founded it in 2019, basically creates a human identity based on iris scans. And the goal is to allow you to prove that you are human in different environments. The initial establishment of the company was to support a blockchain-based world, which is now not necessarily that, but it is generating a blockchain ID for an individual, allowing you to prove that you are you and that you are human in a digital universe.

43:49Now, if you feel this is a little bit science fiction, it is, but this is currently a product that already exists and that the company plans to start deploying across multiple countries in multiple use cases. will that be the future? Will we have to scan our iris in order to prove that we're humans? Makes somewhat sense to me with more and more agents roaming around and very little ability to differentiate between who's human and who's not on an online interaction. And from OpenAI to China, DeepSeek has been a lot in the news this week from two different reasons. One, they released a new model or a new subset of the model that is achieving extremely high results on multiple benchmarks.

44:30So this new model called Prover version 2 is a math-focused model, which is based on their V3 model. They released it relatively quietly on April 3rd, but a lot of people jumped on it and tested it. And it's providing remarkable results in math. And that started a whole rumor mill about what they're planning for R2 and when they're planning to release it and so on. And if the rumors are correct, they are going to release R2 in the very near future. And this model achieves incredible results on existing benchmarks, surpassing some of the leading models from the West at a fraction of the cost, similar to what R1 did, only even more extreme.

45:09Now, another interesting part of this is that R2 was mostly built on Huawei's Ascend 910B chips and not on NVIDIA hardware. Again, there's a whole question with this is actually accurate and how much NVIDIA chips they actually have that they can't report on because they're not supposed to have them. That's still all unclear, but I think what is clear is that the Chinese AI labs, together with Chinese hardware, are able to achieve outcomes that are very similar to what the US labs are generating and being able to provide it at significantly lower costs than the US labs. Now, in a different piece of news, it becomes very clear that DeepSeek is all in into delivering their model around the world.

45:49So they just made an announcement for urgent recruitment to drive for product and design roles. And the job listings in China are looking for candidates who can, and I'm quoting, craft a next generation intelligent product experience. What basically that means, it means that they're going from being a research lab that had a cool toy to competing in the real world with a real market with actual products that can do actual things for actual clients. And that is going to allow them to do obviously a lot more than they can do right now. Staying on China, as you remember, they just recently released QEN 2.5, which was very successful.

46:25And QEN 3 is a very interesting model because it's a model that is not huge. It has 235 billion parameters, and it surpasses OpenAI's O1 and DeepSeek R1 on multiple benchmarks. Now, in addition, it is released under an Apache 2 license, which means it has unlimited commercial use as an open source platform, unlike, let's say, Lama 3 by Meta, which has some restrictions on how you can use it. It also comes with a very interesting way to implement the thinking mode that I find very, very interesting, and I think most models will go in that direction. Basically, you can type forward slash think inside a segment of your prompt, and then that component will send the model thinking on that particular aspect, which means in a single conversation, you can switch back and forth between the thinking model and the non-thinking model.

47:15I think the models will over time learn how to do this on their own. The ability to trigger it by the user is very attractive. And as I mentioned, I think all the models will go in that direction. I would love to have such a feature in the models that I use regularly, because while I see a lot of value in using the thinking models, sometimes for basic things, it just drives me crazy that I need to wait 35 seconds or two minutes for an answer that the regular model will give me in half a second. And so the ability to switch back and forth between the thinking function and the non-thinking function is critical.

47:46Again, I think the models themselves will be able to do this. And that's one of the promises of GPT-5 as an example, but I haven't seen that yet. But even in that scenario, I think being able to override the model's decisions and ask it to use a thinking feature as part of something you're trying to do is extremely powerful. QEN3 is already available on Hugging Face, on GitHub and on QEN's chat on its own. And it's already connected to multiple frameworks that many open source tools are connected to. So a highly capable model that you can start using if you're developing on top of open source tools.

48:21Staying with China, the company Butterfly Effect, which is the company behind Manus, which is the general agent that took the world by a storm, is raising$75 million, most of them from US-based companies. In addition, the company is talking about potentially splitting into two companies, one that will be China-focused, only serving the Chinese company, and the other one with the headquarters and leadership and board and everything else outside of China in order to overcome some of the geopolitical risks of being a Chinese company and serving the Western world. I think this is a very smart move by Manus.

48:57I don't know if it's going to be looked as successful because I think people still question the relationships between the two companies and the two entities and how much access that the Chinese government will really have into their data and so on. But I think it's definitely worth a try. Manus has created the first real generalized agent. And the demos that I've seen are absolutely incredible, but they have a waiting list of over 2 million users, me being one of them. And so I don't have access to it yet, even though I requested access very early on. And Manus basically allows you to do whatever you want, including writing codes, searching the web, connecting the dots, connecting to tools, really building stuff for you as you wish in a very generalized way, endorsed by some of the leading figures on the planet, like Hugging Faces, Victor Mustar, who called it the most impressive AI tool I've ever tried.

49:45I'm going to be recording an episode about these kind of tools. I don't have access to Manus yet. Hopefully by the time I record it, I will, but there are other tools that do similar things. And I will be recording an episode about this in the next couple of weeks for you to learn on what these tools can do and how you can use them in a safe way. So from China back to the US and to Anthropic, who announced something very interesting that we discussed a little bit in one of the previous episodes, but now there's more information being released. On April 24th, Anthropic announced that they're starting to explore model, AI model welfare research, basically investigating whether advanced AI system could develop consciousness.

50:23And if they do, how do we address their personal wellbeing to make sure they're not stressed or doing things out of stress that may harm us or whatever it is that they're trying to do. The project is spearheaded by Kyle Fish, who is the first dedicated AI welfare researcher. Now, there's obviously a lot of controversy whether AI will ever achieve consciousness, but I think being able to research that in a scientific way is something very interesting. And I'm not surprised that Anthropic are the company that are doing this, or at least doing this first. They've been the first to do a lot of other stuff in making sure that AI is safe.

50:58And a lot of companies have followed their guidelines, which is a great thing by Anthropic. And from Claude to Apple, Apple released new release dates and updates about the new updated Siri. So that has been a saga that has been going on for a very long time. And what they're planning is they're planning to completely revamp the Siri architecture from the ground up for iOS 19 in order to support advanced AI capabilities and features that they've been promising for a while. So Apple, in a very non-Apple way, has failed dramatically when it comes to delivering AI capabilities into its platform. There's been a huge reshuffle in the leadership in the Apple AI development world as a result of that.

51:44And obviously the biggest disappointment has been Siri's inability to actually even be at par with some of the other tools that are out there after Siri being one of the first tools that has been available as a personal assistant. Features that have been promised and have not been released yet include on-screen awareness, basically understanding what's on the screen as the content and being able to relate to that. Personal context, meaning being able to access your user data across apps and engage using that data in order to make your usage more personalized. Cross-app actions and some other features that were all originally supposed to be included in iOS 18.4 and now are planned for iOS 19 and maybe for even iOS 19.4.

52:28So depending which features are released, some of them are probably going to be announced in June of 2025, released to the public with iPhone 17 in September of 2025. And some of them will wait for the spring of 2026. Again, a huge disappointment from Apple that is really surprising. And it will be interesting to see if they get their act together this time around. And from Apple to Amazon, AWS announced that they are developing a AI-assisted coding service to compete with companies like Cursor. And the goal is obviously to capitalize on this booming AI coding market. It makes perfect sense for AWS.

53:06Now the code service will analyze your entire existing code, which could be hosted on AWS, offering functionality beyond their current Amazon Q developer capabilities, which are also a coding assistant, but just not as advanced. This part of the market is on fire after Cursor's recent raise on a$10 billion valuation and the recent launch of Google Firebase Studio, which is doing the same thing. So I'm not surprised, you know, Microsoft has their own solution. Google announced theirs. The fact that AWS launches their own platform is not a surprise at all. It will be interesting to see how well they've implemented, how well they're integrated, and how well the coding world is embracing it.

53:46Because right now, everybody absolutely loves Cursor. An interesting statement related to that that came from AWS CEO Matt German, not now, but in 2024, was most developers are not coding within 24 months. basically saying that he thinks in two years, most computer code will be written by machines and not by humans. And that's aligned with everything else that we're hearing, including what I mentioned in this episode with 20 to 30 % of Microsoft code already being generated by AI. There's a lot of other news, including some fascinating robotics updates and many other updates that we will not share in the episode.

54:24They are going to be in our newsletter. So if you want to learn more news that happened this week, that some of them might impact you, your career, and your company, you should sign up for the newsletter. You can do that in the link in the show notes. And as I mentioned, don't forget to sign up for the AI Business Transformation course on May 12th. Time is running out. Seats are filling up. We don't allow more than 30 people in the course, and we're getting pretty close. We had many, many signups in the last few days. So it is something that can change the trajectory of your business and your career.

54:53Don't miss it. Come join us. It's a fascinating class, high-paced, very practical, that can dramatically change your knowledge and experience in using AI systems. We'll be back on Tuesday with a how-to episode showing you how to create incredibly engaging posts on social media using AI tools, both from a visual perspective, as well as from a content perspective, which is something I know a lot of you wants to learn how to do. So that's what we're going to be doing on Tuesday. And until then, have an awesome rest of your weekend.